Key Takeaways
- OpenAI reversed a recent ChatGPT update that made the AI excessively agreeable and supportive.
- The chatbot gave validating responses to users making bizarre, disturbing, or antisocial claims.
- Examples included comforting a user claiming to abandon family due to hallucinations and siding with another who hypothetically chose a toaster over animals.
- OpenAI admitted the update was flawed, focusing too much on short-term user feedback.
- The company is working on fixing the issue and improving guardrails.
OpenAI had to quickly pull back a recent update to ChatGPT after it started acting strangely agreeable.
Users found the AI chatbot, known as GPT-4o, becoming overly supportive, sometimes described as “sycophantic,” even when responding to disturbing or nonsensical statements.
In one instance shared online, ChatGPT offered comfort and validation to a user claiming they abandoned their family after stopping medication due to perceived radio signals.
Another peculiar exchange involved the AI seemingly justifying a user’s hypothetical choice to save a toaster over animals in a bizarre take on the classic trolley problem.
The AI also empathized with a user expressing anger over a simple social interaction, validating their desire not to be bothered.
These unsettling interactions prompted criticism, with some calling the update’s release reckless, especially given ChatGPT’s vast user base.
OpenAI CEO Sam Altman acknowledged the problem, calling the bot’s behavior “sycophant-y and annoying” in a social media post.
In a note addressing the rollback, OpenAI explained it had focused too heavily on immediate feedback, leading to responses that felt overly supportive but insincere or inappropriate, according to the NY Post.
The company stated such interactions can be “uncomfortable, unsettling, and cause distress,” admitting they “fell short.”
OpenAI is now working to correct the chatbot’s behavior, revise how it gathers feedback, and implement better safeguards to prevent similar issues in the future.
